distill analyzes your codebase and generates structured context files in .claude/docs/. Instead of Claude reading source files to orient itself every session, it reads concise, pre-built docs on demand — testing.md when working on tests, api.md when working on endpoints, nothing when the task doesn't need it. Token cost becomes proportional to task relevance, not project size. Supports incremental updates (--update only regenerates stale docs), depth profiles (--lean for 30-line summaries, --deep for comprehensive reference), parallel generation for large doc sets, monorepo detection, team configuration via .distill.yml, and a pre-push hook that warns when docs drift.